Re: Balloon driver bug in increase_reservation

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Wed, Sep 04, 2013 at 04:42:16PM +0100, Wei Liu wrote:
> On Wed, Sep 04, 2013 at 11:14:30AM -0400, Konrad Rzeszutek Wilk wrote:
> [...]
> > > > 
> > > > Otherwise could simply implement a is_balloon_scratch_page function that
> > > > checks whether a given pfn corresponds to any of the scratch pages (it
> > > > doesn't need to be the scratch page of this cpu).
> > > 
> > > That's quite expensive IMHO, especially when you have lots of CPU's and
> > > lots of ballooned pages.
> > 
> > Fortunatly you don't have to take lock. The PFNs for the scratch pages are
> > set in stone for each vCPU and don't change (unless the CPU goes down, but
> > then the 'for_each_online_cpu' would omit said CPU).
> > 
> > And I think the balloon driver does everything from one workqueue so
> > the check can done there?
> 
> Well, what are the chances that you have 256 CPUs and then need to
> balloon 2K pages (only 8MB)... andd what's the frequency you need to do
> that... Maybe I'm just paranoid to imagine all those extreme use cases.

Fortunatly it is a slow process. It does not have to happen immediately so
we can also take a lock if need to. Just as long as the lock is not taken
in the M2P or P2M code I think we are fine.

> 
> Wei.
> 
> > > 
> > > Wei.
> > > 
> > > _______________________________________________
> > > Xen-devel mailing list
> > > Xen-devel@lists.xen.org
> > > http://lists.xen.org/xen-devel